Microblog Analysis from the viewpoint of Social Sensor

摘要

Twitter users make tweets related to their emotion and description for local events around them. We can get such tweets via timeline and know informations of local events. In this paper, we regard Twitter users as social sensors and tweets as the outputs of the social sensors. We aim to find local events from Twitter by natulal language proseccing. We develop a weather sensing system using Twitter. First, we use Twitter search API to gather raw tweets which have a word related to snowing. And then we classify tweets by Support Vector Machine into the tweets related to snowing and irrelated tweets. Before the step of class-fication, we have to convert tweet into vector representation. We make feature vectors by N-gram model and then we could make a classifier which achieves high recall rate. At last, we estimate where the event is happened by keyword matching and make time series chart. Carrying out evaluation experiments, we can estimate where it is snowing in Japan and discover a characteristic of Twitter as a social sensor system. Each users as social sensors had sensitivity for each events according to loation.